AI Models Crack WWII Enigma Codes, Defying Cryptanalysts

Overview

A pair of advanced language models from OpenAI and Anthropic have successfully cracked long-unsolved Enigma codes from World War II, marking a significant milestone in the application of artificial intelligence to historical cryptanalysis. The breakthrough challenges traditional methods and suggests that modern LLMs can perform complex archival research and logical deduction at speeds previously impossible for human analysts.

Cryptanalysts Carter Leffen and Jack Willis independently utilized OpenAI’s Astra and Anthropic’s Opus 5 models to decode messages that have baffled researchers for decades. Their achievements highlight a new era where AI agents do not merely assist in data processing but actively engage in independent investigative work, simulating historical machines and navigating vast archives to find solutions.

The Astra Breakthrough

Carter Leffen, a developer with a background in cryptography, approached the problem by instructing OpenAI’s Astra model to locate and decode an unbroken message from a public database of Enigma communications. The task was daunting; the machine’s encryption logic was notoriously complex, and many historical records suffered from transcription errors or incomplete metadata.

Rather than relying solely on pattern matching, the AI conducted its own archival research. It identified context clues within the message structure and, crucially, built a functional simulator of the Enigma machine. This self-directed approach allowed the model to test various configurations and recover the plaintext of a message that had remained encrypted since 2005.

“GPT–6 Astra is behaving like a very professional cryptanalyst and archive researcher. What it has achieved in two days would take a human researcher weeks or even months.”

The model’s capabilities extended beyond simple decoding. Leffen utilized Astra to construct an interactive website explaining the cryptographic problem and its solution, demonstrating the AI’s ability to synthesize technical findings into accessible educational content. This holistic workflow—from research to simulation to dissemination—illustrates the evolving role of AI as a comprehensive analytical tool.

Validation and Archival Mysteries

The significance of Leffen’s discovery was reinforced by Frode Weirerud, a retired electrical engineer and cryptologist who maintains the Crypto Cellar database. Weirerud spent several weeks verifying the Bundesarchiv files referenced by the AI, ultimately confirming the validity of the decryption. His validation has left him in what he describes as a state of awe, noting that the AI’s performance surpassed his own extensive manual efforts.

However, the investigation uncovered a puzzling element regarding the sources of information accessed during the process. Logs from the Astra model include references to archived messages held in a “private collection” that are not part of the public Crypto Cellar database. Weirerud remains uncertain whether the model accessed these restricted materials or if it inferred their existence from scattered discussions online.

Speculation suggests the AI may have encountered data shared by other researchers on various forums or potentially accessed digitized records from German government archives. This raises important questions about the boundaries of AI data retrieval and the potential for models to encounter information that is not widely indexed or publicly acknowledged.

The Second Breakthrough

On September 21, another significant development emerged when cryptanalyst Jack Willis announced that he had used Anthropic’s Claude Opus 5 to break a different unsolved Enigma message. Unlike Leffen’s relatively open-ended prompt, Willis provided the AI with specific guidance, focusing on the known signature of a particular officer’s name embedded within the ciphertext.

This targeted approach allowed Opus 5 to leverage pattern recognition effectively. By isolating the unique characteristics of the officer’s signature, the model was able to deduce the underlying encryption settings and recover the original text. Willis’s success underscores the versatility of current AI models, which can adapt to both broad investigative tasks and highly specific, guided challenges.

The collaboration between human experts and AI systems appears to be yielding results faster than either could achieve independently. Human analysts provide the historical context and validation frameworks, while the models offer the computational power to test millions of permutations and synthesize disparate pieces of information.

Implications for Cryptanalysis

Weirerud notes that only seven unbroken Enigma messages remain in existence, along with one message where the plaintext is known but the specific coding configuration is still unbroken. Given the pace of recent developments, it is highly probable that these final barriers will fall in the near future.

This rapid progress suggests a paradigm shift in how historical cryptographic challenges are approached. Traditional cryptanalysis relied heavily on manual computation and the intuition of human experts working in isolation. Today, AI agents can operate as autonomous researchers, capable of simulating machinery, accessing global archives, and cross-referencing data points in real-time.

The success of Astra and Opus 5 in solving these wartime puzzles serves as a powerful proof of concept. It demonstrates that large language models possess the reasoning capabilities necessary to handle complex, multi-step problems that require both logical rigor and creative insight. As these technologies continue to evolve, they promise to unlock new insights from history’s most guarded secrets, including the enigmatic codes of WWII.